Announcement

Collapse
No announcement yet.

CPUFreq Scaling Tests With AMD's Kaveri On Linux 3.16

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • CPUFreq Scaling Tests With AMD's Kaveri On Linux 3.16

    Phoronix: CPUFreq Scaling Tests With AMD's Kaveri On Linux 3.16

    In past months on Phoronix we have done many CPUfreq tests to see the impact of the different scaling governors and with Intel hardware how the CPUfreq driver compares to Intel P-State in terms of performance and power consumption. As we haven't done such ACPI CPUfreq tests on the AMD side in a long time, here's some benchmarks using a Kaveri APU and selecting the different scaling governors during performance benchmarking.

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    This looks OK, but someone needs to try racing games with ondemand vs performance and tell a difference .

    Supertuxkart for example should do a thing .

    Comment


    • #3
      Maybe bullet physics or physics overall make is nearly halfed performant and sttuters with ondemand

      Comment


      • #4
        Why is it that the frequency is not locked to the maximum in performance mode? Isn?t it what this governor does anymore?

        Comment


        • #5
          Originally posted by stqn View Post
          Why is it that the frequency is not locked to the maximum in performance mode? Isn?t it what this governor does anymore?
          does it use turbo?
          how are frequencies actually measured? is it by core or average over all cores? in which cases were all of the cores used? when did turbo freq kick in? is there a way to find out? does it actually mean that for the graphics benchmarks the GPU or GPUtoRAM bandwidth was the bottleneck?
          what can we learn from this?
          is the memory also clocked up/down by OS or is it done within the hardware itself?

          Comment


          • #6
            Originally posted by stqn View Post
            Why is it that the frequency is not locked to the maximum in performance mode? Isn?t it what this governor does anymore?
            Yeah booting with processor.max_cstate=0 show some difference so it is not right at maximum in performance mode .

            Comment


            • #7
              Originally posted by dungeon View Post
              Yeah booting with processor.max_cstate=0 show some difference so it is not right at maximum in performance mode .
              Actually performance mode is not at quilty it is the same if i not compile so not use CPU Frequency scaling at all . So performance mode is fine .

              Only max_cstate=0 is very slightly faster

              Comment


              • #8
                If you want to see the real CPU frequency, use cpufreq-aperf (must load the msr module beforehand).

                Comment


                • #9
                  Interesting article, but I still prefer PState with NOOP for my desktop PC.

                  Then again, I do have an old netbook laying around...

                  Comment

                  Working...
                  X